AI router: one endpoint, the right LLM for every request

One API in front of many models

Your code calls one endpoint with one key. The router translates the request for whichever provider it picks and returns the answer in the same shape, so changing models is a setting, not a rewrite.

A decision per request

The router picks the model for this request: a strong coder for a bug fix, an inexpensive model for bulk tagging, a fast one for a one-line reply. A smart router makes that choice automatically, where a hard-coded model name never changes.

Fallbacks when a provider fails

If the chosen model errors or times out, the router tries the next one, at another provider where it can, so an outage at one provider need not become an outage in your product.

One place for logs and spend

Every request passes one point, so the model, latency, tokens and cost of each one land in one log, with budgets and limits on top.

Rules you write

Requests that match a pattern go to a model you named: code to one, long documents to another. Predictable, and as current as the last time someone revisited the rules.

Price and speed

The least expensive or the fastest model or provider that can serve the request. Clear on cost and latency, and quality is left to you.

A prompt classifier

A small model labels each prompt with a task type and maps the type to a model, from the router’s own training data or from what other users choose for that type.

Benchmark results

Every model runs the same tasks on a schedule, and the request goes to the model placing highest on the board for that job. The basis is public and checkable, and it moves when a model changes. This is how the Smart Router decides, with response time, price and your own traffic split available as modes of their own.

As of 30 Oct 2026, 15:27 UTC · the router’s own picks for an example caller with keys for all 6 providers

Modes that name a job

auto-coding, auto-reasoning, auto-tooling and auto-best rank on the boards. auto-value and auto-fastest-quality choose among the models within 5 points of the best, then by cost or speed. auto-task reads each request and picks the board itself. A statistical tie goes to your tie-break, cheapest per task by default.

Drift-aware

A model whose drift detector has fired on the suite your mode relies on, or that has an open canary incident, moves behind the healthy ones. The routing reason names any model passed over.

Fallbacks and a circuit breaker

Up to 5 fallbacks on every request, your own order first. A model that fails 5 times in a row is skipped for 60 seconds and probed before it returns.

Your keys, no markup

Requests go out with your own provider keys, stored with AES-256-GCM, and each provider bills you at its own list price. Plans count routed requests: 1,000 a month on Free.

Two APIs

An OpenAI-compatible /v1/chat/completions with streaming, tool calls and JSON output, and a native Anthropic /v1/messages for Claude Code.

Explained on every response

The X-AISM-Reasoning header names the model, the provider and why it won.

1. Add your provider keys

Create a free account and add a key for each provider you use. Keys are stored encrypted with AES-256-GCM, and the modes rank only the models you hold keys for.

2. Create a router key

One aism_ key reaches every provider you connected and every mode. It goes wherever your code expects an OpenAI key.

3. Change the base URL

Point the client at https://aistupidlevel.info/v1 and send a mode such as auto-coding as the model.

from openai import OpenAI

client = OpenAI(
    base_url="https://aistupidlevel.info/v1",
    api_key="aism_your_key_here",
)

response = client.chat.completions.create(
    model="auto-coding",
    messages=[{"role": "user", "content": "Fix the failing test in parser.py"}],
)
print(response.choices[0].message.content)

• AI aggregator
• Best AI routers
• OpenRouter alternatives
• Smart Router